Space-Time Adaptive Processing Based on Interpretable Multimodule Convolutional Neural Network

Abstract

Space-time adaptive processing (STAP) performance in complex clutter environments often degrades due to the difficulty in obtaining sufficient independent and identically distributed (IID) training samples. Sparse recovery (SR) STAP reduces IID sample requirements but faces challenges in parameter tuning and computational complexity. Deep-learning (DL) STAP methods also lower IID sample needs and reduce online computation, but their poor interpretability limits reliability. In addition, both SR and DL STAP methods are prone to off-grid issues. To address these challenges, this article proposes a multimodule deep convolutional neural network that combines data-driven and model-driven approaches to achieve fast and accurate clutter covariance matrix estimation under small-sample conditions. The network comprises four parts: a channel self-attention module, data modules, prior modules, and a hyperparameter module. Each module has a clear mathematical foundation and physical significance, enhancing the interpretability of the network. Meanwhile, the network leverages prior knowledge of clutter ridges to nonuniformly partition the spatial-Doppler profile, effectively mitigating the impact of off-grid issues. Both simulated and measured data demonstrate that the proposed method outperforms existing small-sample STAP methods in clutter suppression within nonhomogeneous clutter environments, significantly reducing computational time.